Tencent Open-Sources WeMM-Embedding Multimodal Models, Topping MMEB Leaderboards

Tencent open-sourced the WeMM-Embedding family on Hugging Face — 2B, 4B and 9B variants, all built on Qwen3.5 under Apache 2.0. The models embed text, images, video and visual documents into a single vector space, supporting cross-modal retrieval with a focus on mixed-media, video-corpus scenarios. The 9B model tops both MMEB-v2 and MMEB-v3, and the 2B variant beats 8B-class competitors with roughly a quarter of the parameters.

Confirmed

Three sizes: WeMM-Embedding-2B/4B/9B, based on Qwen3.5, Apache 2.0
Multimodal document representation: a single input can interleave multiple images or videos
MMEB-v2 (78 datasets): 9B scores 80.6, ahead of closed-source DME-Medium (78.4) and Qwen3-VL-Embedding-8B; the 2B model outperforms the 8B model with about a quarter of the parameters
Largest gains on the video subset: WeMM-9B 74.3, WeMM-2B 70.8, Qwen3-VL-Embedding-8B 67.1 — the 2B leads the 8B by 3.7 points
MMEB-v3 (190 tasks spanning text, agents, cross-modal retrieval): 9B tops at 59.5, 4B at 58.2, 2B beats 7B/8B-class models

Why it matters

Open and permissively licensed, offering a ready-to-deploy unified embedding solution for cross-modal retrieval, RAG and video search
Small models beating much larger competitors on authoritative benchmarks signal efficiency advantages in architecture and training, lowering deployment costs

2026-09-02 ~ 2026-09-02 · 6 related posts

Primary sources

Tencent open-sources WeMM-Embedding: unified text/image/video embeddings, Apache 2.0 — tomaarsen · 2026-09-02 21:35 (UTC+08:00)
Tencent releases WeMM-Embedding multimodal models for mixed text, image, and video retrieval — tomaarsen · 2026-09-02 21:35 (UTC+08:00)
WeMM-Embedding-2B edges out Qwen3-VL-Embedding-8B on MMEB-v2 at quarter size — tomaarsen · 2026-09-02 21:35 (UTC+08:00)

[source] Tencent open-sources WeMM-Embedding: unified text/image/video embeddings, Apache 2.0 — tomaarsen · 2026-09-02
WeMM-Embedding tops MMEB-v2: its 2B model edges out 8B rivals at a quarter of the size — tomaarsen · 2026-09-02
Video is the widest gap: WeMM-Embedding-2B beats an 8B by 3.7 points — tomaarsen · 2026-09-02
[source] Tencent releases WeMM-Embedding multimodal models for mixed text, image, and video retrieval — tomaarsen · 2026-09-02
WeMM-Embedding tops MMEB-v3: 9B scores 59.5, 2B beats every 7B/8B model — tomaarsen · 2026-09-02

1 near-duplicate retellings: tomaarsen